The sound is comparable to DNA, which is infinitely useful under the blessing of AI 【Full Text】

If the security news technology news let you simply listen to a person's voice, what information can you hear? age? gender? Hometown? For AI, these are too difficult. Recently, a study by the Massachusetts Institute of Technology showed that the trained AI can not only get information about people's gender, race, age, etc. from the voice, it can even hear what you look like!

It is understood that this AI that can complete the operation of "listening to distinguish people" mainly depends on a neural network model called Speech2Face to complete training. The model is divided into two parts. One is a speech encoder, which is responsible for facial feature analysis and prediction of the input speech; the other is a face decoder, which integrates and generates the input facial features.

In actual operation, the researchers put a dataset of one million video clips into the model, and then let the AI ​​perform self-training for a period of time. Then, with only 6 seconds of speech, the AI ​​can achieve human The collection and restoration of facial features, and present a good image.

From the part of the training results given by the MIT research team, we can find that Speech2Face can better identify the gender, and can also be better distinguished between Caucasians and Asians, and also for the 30-40 and 70-year-olds The sound hit rate is slightly higher. However, because AI's "hearing" is not 100% reliable, and the training materials are not rich enough, it will also produce a lot of recognition errors, and the ability to distinguish black voices is also weak.

Although the technology is not yet perfect, it is more than enough to meet the initial MIT vision. The research team pointed out that they did not train the AI ​​function to accurately restore the speaker's appearance, but only to study the relationship between speech and appearance, and use this to generate various cute cartoon user avatars.

You may think that such a technology is a bit overkill for avatar generation, don't worry! Because of the similar technology, other research institutions are also actively developing, and some have already invested in some meaningful application scenarios.

For example, Carnegie Mellon University has published a similar study that can guess the speaker's age, height, weight, space and environmental information from the voice. Researchers at the university believe that sound is like human DNA, which contains rich and unique information and can be used in all walks of life.

When the accuracy of the technology identification and restoration exceeded 60%, they began to formally put into the society for application testing. At present, the United States Coast Guard is still using this technology to identify malicious alarmers. This technology helps them distinguish whether the alarmers are pranks and narrows the scope of the investigation. This allows them to reduce nearly 150 prank calls every year. , Saving a lot of police resources.

It is understood that Carnegie Mellon University's research team eventually conceived of using AI, a "listening to people" technology to remotely diagnose Parkinson's disease. I hope this technology can open the door to modern medical innovation and provide solutions and solutions for intractable diseases and some terminal illnesses.

In addition to using similar technologies for criminal investigation and medical treatment, in reality, the same technology has also been applied to many scenarios and fields such as banking, insurance, customer service, and recruitment. Among them, banks such as HSBC and Morgan use voiceprint recognition to protect user account security; Metropolitan Manual Insurance Company uses AI systems to identify customer emotions and feelings; some insurance companies use this technology to determine the intention of the caller; also Some companies use the technology for recruitment ...

In addition, in 2017, Toyota Motor also applied the technology to driving at the CES conference. AI is loaded on the camera, sensor, and car voice system to help determine whether the driver is in a fatigued driving state, and promptly remind. This technology makes the driver's driving more intelligently guaranteed.

In short, no matter what kind of application, the function of AI "listening to people" is undoubtedly of great value. We have reason to believe that this technology will eventually appear more and more in future life and production. However, if AI wants to truly become a good helper and partner for people in the future, it needs further upgrades and breakthroughs, and the road to development needs to be looked forward to!